1 //===- UseListOrder.cpp - Implement Use List Order functions --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Implement use list order functions to modify use-list order and verify it
11 // doesn't change after serialization.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/IR/UseListOrder.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/AsmParser/Parser.h"
20 #include "llvm/Bitcode/ReaderWriter.h"
21 #include "llvm/IR/LLVMContext.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/FileSystem.h"
26 #include "llvm/Support/FileUtilities.h"
27 #include "llvm/Support/MemoryBuffer.h"
28 #include "llvm/Support/SourceMgr.h"
33 #define DEBUG_TYPE "use-list-order"
37 static cl::opt<bool> PreserveBitcodeUseListOrder(
38 "preserve-bc-use-list-order",
39 cl::desc("Experimental support to preserve bitcode use-list order."),
40 cl::init(false), cl::Hidden);
42 bool llvm::shouldPreserveBitcodeUseListOrder() {
43 return PreserveBitcodeUseListOrder;
46 static void shuffleValueUseLists(Value *V, std::minstd_rand0 &Gen,
47 DenseSet<Value *> &Seen) {
48 if (!Seen.insert(V).second)
51 if (auto *C = dyn_cast<Constant>(V))
52 if (!isa<GlobalValue>(C))
53 for (Value *Op : C->operands())
54 shuffleValueUseLists(Op, Gen, Seen);
56 if (V->use_empty() || std::next(V->use_begin()) == V->use_end())
57 // Nothing to shuffle for 0 or 1 users.
60 // Generate random numbers between 10 and 99, which will line up nicely in
61 // debug output. We're not worried about collisons here.
62 DEBUG(dbgs() << "V = "; V->dump());
63 std::uniform_int_distribution<short> Dist(10, 99);
64 SmallDenseMap<const Use *, short, 16> Order;
65 for (const Use &U : V->uses()) {
68 DEBUG(dbgs() << " - order: " << I << ", U = "; U.getUser()->dump());
71 DEBUG(dbgs() << " => shuffle\n");
73 [&Order](const Use &L, const Use &R) { return Order[&L] < Order[&R]; });
76 for (const Use &U : V->uses())
77 DEBUG(dbgs() << " - order: " << Order.lookup(&U) << ", U = ";
82 void llvm::shuffleUseLists(Module &M, unsigned SeedOffset) {
83 DEBUG(dbgs() << "*** shuffle-use-lists ***\n");
84 std::minstd_rand0 Gen(std::minstd_rand0::default_seed + SeedOffset);
85 DenseSet<Value *> Seen;
87 // Shuffle the use-list of each value that would be serialized to an IR file
88 // (bitcode or assembly).
89 auto shuffle = [&](Value *V) { shuffleValueUseLists(V, Gen, Seen); };
92 for (GlobalVariable &G : M.globals())
94 for (GlobalAlias &A : M.aliases())
99 // Constants used by globals.
100 for (GlobalVariable &G : M.globals())
101 if (G.hasInitializer())
102 shuffle(G.getInitializer());
103 for (GlobalAlias &A : M.aliases())
104 shuffle(A.getAliasee());
105 for (Function &F : M)
106 if (F.hasPrefixData())
107 shuffle(F.getPrefixData());
110 for (Function &F : M) {
111 for (Argument &A : F.args())
113 for (BasicBlock &BB : F)
115 for (BasicBlock &BB : F)
116 for (Instruction &I : BB)
119 // Constants used by instructions.
120 for (BasicBlock &BB : F)
121 for (Instruction &I : BB)
122 for (Value *Op : I.operands())
123 if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
128 DEBUG(dbgs() << "\n");
134 std::string Filename;
136 bool init(const std::string &Ext);
137 bool writeBitcode(const Module &M) const;
138 bool writeAssembly(const Module &M) const;
139 std::unique_ptr<Module> readBitcode(LLVMContext &Context) const;
140 std::unique_ptr<Module> readAssembly(LLVMContext &Context) const;
143 struct ValueMapping {
144 DenseMap<const Value *, unsigned> IDs;
145 std::vector<const Value *> Values;
147 /// \brief Construct a value mapping for module.
149 /// Creates mapping from every value in \c M to an ID. This mapping includes
150 /// un-referencable values.
152 /// Every \a Value that gets serialized in some way should be represented
153 /// here. The order needs to be deterministic, but it's unnecessary to match
154 /// the value-ids in the bitcode writer.
156 /// All constants that are referenced by other values are included in the
157 /// mapping, but others -- which wouldn't be serialized -- are not.
158 ValueMapping(const Module &M);
160 /// \brief Map a value.
162 /// Maps a value. If it's a constant, maps all of its operands first.
163 void map(const Value *V);
164 unsigned lookup(const Value *V) const { return IDs.lookup(V); }
169 bool TempFile::init(const std::string &Ext) {
170 SmallVector<char, 64> Vector;
171 DEBUG(dbgs() << " - create-temp-file\n");
172 if (auto EC = sys::fs::createTemporaryFile("use-list-order", Ext, Vector)) {
174 DEBUG(dbgs() << "error: " << EC.message() << "\n");
177 assert(!Vector.empty());
179 Filename.assign(Vector.data(), Vector.data() + Vector.size());
180 Remover.setFile(Filename);
181 DEBUG(dbgs() << " - filename = " << Filename << "\n");
185 bool TempFile::writeBitcode(const Module &M) const {
186 DEBUG(dbgs() << " - write bitcode\n");
187 std::string ErrorInfo;
188 raw_fd_ostream OS(Filename.c_str(), ErrorInfo, sys::fs::F_None);
189 if (!ErrorInfo.empty()) {
190 DEBUG(dbgs() << "error: " << ErrorInfo << "\n");
194 WriteBitcodeToFile(&M, OS);
198 bool TempFile::writeAssembly(const Module &M) const {
199 DEBUG(dbgs() << " - write assembly\n");
200 std::string ErrorInfo;
201 raw_fd_ostream OS(Filename.c_str(), ErrorInfo, sys::fs::F_Text);
202 if (!ErrorInfo.empty()) {
203 DEBUG(dbgs() << "error: " << ErrorInfo << "\n");
211 std::unique_ptr<Module> TempFile::readBitcode(LLVMContext &Context) const {
212 DEBUG(dbgs() << " - read bitcode\n");
213 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOr =
214 MemoryBuffer::getFile(Filename);
216 DEBUG(dbgs() << "error: " << BufferOr.getError().message() << "\n");
220 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOr.get());
221 ErrorOr<Module *> ModuleOr = parseBitcodeFile(Buffer.release(), Context);
223 DEBUG(dbgs() << "error: " << ModuleOr.getError().message() << "\n");
226 return std::unique_ptr<Module>(ModuleOr.get());
229 std::unique_ptr<Module> TempFile::readAssembly(LLVMContext &Context) const {
230 DEBUG(dbgs() << " - read assembly\n");
232 std::unique_ptr<Module> M(ParseAssemblyFile(Filename, Err, Context));
234 DEBUG(dbgs() << "error: "; Err.print("verify-use-list-order", dbgs()));
238 ValueMapping::ValueMapping(const Module &M) {
239 // Every value should be mapped, including things like void instructions and
240 // basic blocks that are kept out of the ValueEnumerator.
242 // The current mapping order makes it easier to debug the tables. It happens
243 // to be similar to the ID mapping when writing ValueEnumerator, but they
244 // aren't (and needn't be) in sync.
247 for (const GlobalVariable &G : M.globals())
249 for (const GlobalAlias &A : M.aliases())
251 for (const Function &F : M)
254 // Constants used by globals.
255 for (const GlobalVariable &G : M.globals())
256 if (G.hasInitializer())
257 map(G.getInitializer());
258 for (const GlobalAlias &A : M.aliases())
260 for (const Function &F : M)
261 if (F.hasPrefixData())
262 map(F.getPrefixData());
265 for (const Function &F : M) {
266 for (const Argument &A : F.args())
268 for (const BasicBlock &BB : F)
270 for (const BasicBlock &BB : F)
271 for (const Instruction &I : BB)
274 // Constants used by instructions.
275 for (const BasicBlock &BB : F)
276 for (const Instruction &I : BB)
277 for (const Value *Op : I.operands())
278 if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
284 void ValueMapping::map(const Value *V) {
288 if (auto *C = dyn_cast<Constant>(V))
289 if (!isa<GlobalValue>(C))
290 for (const Value *Op : C->operands())
294 IDs[V] = Values.size();
298 static void dumpMapping(const ValueMapping &VM) {
299 dbgs() << "value-mapping (size = " << VM.Values.size() << "):\n";
300 for (unsigned I = 0, E = VM.Values.size(); I != E; ++I) {
301 dbgs() << " - id = " << I << ", value = ";
302 VM.Values[I]->dump();
306 static void debugValue(const ValueMapping &M, unsigned I, StringRef Desc) {
307 const Value *V = M.Values[I];
308 dbgs() << " - " << Desc << " value = ";
310 for (const Use &U : V->uses()) {
311 dbgs() << " => use: op = " << U.getOperandNo()
312 << ", user-id = " << M.IDs.lookup(U.getUser()) << ", user = ";
317 static void debugUserMismatch(const ValueMapping &L, const ValueMapping &R,
319 dbgs() << " - fail: user mismatch: ID = " << I << "\n";
320 debugValue(L, I, "LHS");
321 debugValue(R, I, "RHS");
329 static void debugSizeMismatch(const ValueMapping &L, const ValueMapping &R) {
330 dbgs() << " - fail: map size: " << L.Values.size()
331 << " != " << R.Values.size() << "\n";
339 static bool matches(const ValueMapping &LM, const ValueMapping &RM) {
340 DEBUG(dbgs() << "compare value maps\n");
341 if (LM.Values.size() != RM.Values.size()) {
342 DEBUG(debugSizeMismatch(LM, RM));
346 // This mapping doesn't include dangling constant users, since those don't
347 // get serialized. However, checking if users are constant and calling
348 // isConstantUsed() on every one is very expensive. Instead, just check if
349 // the user is mapped.
350 auto skipUnmappedUsers =
351 [&](Value::const_use_iterator &U, Value::const_use_iterator E,
352 const ValueMapping &M) {
353 while (U != E && !M.lookup(U->getUser()))
357 // Iterate through all values, and check that both mappings have the same
359 for (unsigned I = 0, E = LM.Values.size(); I != E; ++I) {
360 const Value *L = LM.Values[I];
361 const Value *R = RM.Values[I];
362 auto LU = L->use_begin(), LE = L->use_end();
363 auto RU = R->use_begin(), RE = R->use_end();
364 skipUnmappedUsers(LU, LE, LM);
365 skipUnmappedUsers(RU, RE, RM);
369 DEBUG(debugUserMismatch(LM, RM, I));
372 if (LM.lookup(LU->getUser()) != RM.lookup(RU->getUser())) {
373 DEBUG(debugUserMismatch(LM, RM, I));
376 if (LU->getOperandNo() != RU->getOperandNo()) {
377 DEBUG(debugUserMismatch(LM, RM, I));
380 skipUnmappedUsers(++LU, LE, LM);
381 skipUnmappedUsers(++RU, RE, RM);
384 DEBUG(debugUserMismatch(LM, RM, I));
392 bool llvm::verifyBitcodeUseListOrder(const Module &M) {
393 DEBUG(dbgs() << "*** verify-use-list-order: bitcode ***\n");
398 if (F.writeBitcode(M))
402 std::unique_ptr<Module> OtherM = F.readBitcode(Context);
406 return matches(ValueMapping(M), ValueMapping(*OtherM));
409 bool llvm::verifyAssemblyUseListOrder(const Module &M) {
410 DEBUG(dbgs() << "*** verify-use-list-order: assembly ***\n");
415 if (F.writeAssembly(M))
419 std::unique_ptr<Module> OtherM = F.readAssembly(Context);
423 return matches(ValueMapping(M), ValueMapping(*OtherM));